Inspiration
Every small business owner has had the same thought: who should we team up with? A coffee shop wonders about a bakery. A skincare brand wonders about a gym. Most of the time the answer is a guess, and a bad guess costs real money: booth fees, inventory, a day of staff time.
The tools that exist make it worse. They answer with confidence, always, even when they shouldn't. We asked a different question: what if the tool recommending a partner tried to talk you out of it before you spent a dollar, and what if it was willing to say "no"?
What it does
Vyuha is a partner finder that argues against itself.
You describe your business and what you want to try, such as a one-day pop-up. Vyuha then:
- Asks Qloo which businesses in other categories share your customers' tastes. These are partners a keyword search would never suggest.
- Checks the map to confirm each candidate is a real, local place.
- Plays devil's advocate. A critic step hunts for reasons each candidate could fail.
- Gives you a verdict in plain English: the best reason to try this partner, the biggest risk, and the cheapest first test to run.
Every claim on screen is tagged with where it came from: Qloo, map data, AI interpretation, a rule-based decision, or unknown. A "How Vyuha decided" page shows each step of the pipeline.
And when the evidence isn't good enough, Vyuha refuses to guess. "No safe partner found" is a real answer, with a specific reason and next steps, not an error.
Two real results
Blue Bottle Coffee, San Francisco, looking for a one-day pop-up partner. Vyuha checked 5 candidates, and the devil's advocate rejected 3 of them. It recommended testing Galleria Park Hotel as a shared-space host, and stated its own biggest risk up front: the taste signal comes from Qloo and is not corroborated by an independent audience source. The first step it suggests costs nothing: verify the hotel's identity and ask about availability and fees before committing.
"Qvxxqz Blorpblorp", Boise, a business that does not exist. Vyuha found no verified listing, invented no partner, and answered no candidate supported, with the reason. This is the behaviour the whole project is built around.
How Qloo makes this possible
Qloo is the discovery engine, not decoration.
- Qloo Search resolves your business to a Qloo entity.
- Qloo Insights returns cross-category entities whose taste signals relate to your business. Those are our candidate partners.
- Candidates without a positive numeric affinity are excluded, and other branches of your own brand are excluded so Vyuha never recommends that you partner with yourself.
Without Qloo, Vyuha would just recommend the obvious competitor next door. Qloo is what finds the unexpected match: the hotel lobby for the coffee brand.
How we built it
- Qloo API (Search and Insights) for cross-category taste discovery
- LangGraph pipeline with separate stages: identity, discovery, verification, comparison and critique, then a rule-based decision, experiment plan, and outreach draft
- OpenStreetMap / Nominatim to verify businesses are real and local
- Groq as the primary model, with a second Groq model as an automatic fallback: on a rate limit or timeout, the same call fails over immediately, and every response is validated against a schema before it is used
- FastAPI backend on Render, Next.js frontend on Vercel, streaming progress to the browser
- A full run uses just two batched AI generations (candidate interpretation and the critic). The decision, experiment plan, and outreach draft are deterministic code, so the final verdict never depends on an AI's mood.
- Dated saved results for every demo, always labelled "Saved result, not live", so a demo never dead-ends if a live service is slow
Challenges we ran into
- Honesty is harder than confidence. The easy version of this project always returns an answer. Building abstention, source tagging, and a critic that can overrule the recommendation took most of our effort.
- We tested like skeptics. When we tried a famous brand, Vyuha briefly suggested another branch of the same brand as its "partner". We fixed the matching so it can't recommend a business to itself. We also found a warning that was showing on every result, which taught users to ignore it, so it now appears only when a real check failed, and it names what failed.
- We audited our own code for fake data. We searched the whole repository for hardcoded or mock results, removed a fabricated step count and stale metrics, and made sure a failed call can never be shown as an empty result.
- Free-tier limits. We hit LLM rate limits mid-run and a backend that sleeps when idle. That led to immediate model failover, a wake-up message, and saved results as a safety net.
Accomplishments we're proud of
- A deployed, working app where the most interesting outputs are the ones that say no
- Every claim traceable to its source, with AI interpretation kept separate from Qloo and map evidence
- Reliable under pressure: model failover, cold-start handling, and clearly dated saved results
- A test suite that was passing at the time of writing: 88 backend tests and 64 frontend tests, plus lint, type checks, and a production build
What we learned
A recommendation is only as useful as your ability to check it. People trust a tool more when it shows its doubts. And "I don't know" is a feature when the cost of being wrong is someone's rent money.
Who it's for, and how it could make money
Independent cafes, shops, and small brands planning collaborations or pop-ups. They have no research team, and a wasted partnership hurts. Possible models: a monthly subscription for small businesses, and per-report pricing for agencies and shopping-district associations that plan collaborations for many members.
Limits, stated plainly
Qloo taste affinity is not proof of shared customers or purchase intent, a map listing does not prove ownership or availability, and no partner is guaranteed to succeed. Results also depend on how much Qloo data exists for a brand and city, and Vyuha says so when it can't find a safe partner. Vyuha has not yet been validated against real partnership outcomes; it recommends small, cheap tests, not predictions.
What's next
- Let owners log how a test went, so rankings improve over time
- Backtest against well-known real collaborations and publish the hits and the misses
- Clearer coverage messages per city and category
- A one-click shareable one-page plan
Try it: https://vyuha-rust.vercel.app/ (pick Blue Bottle Coffee for a recommendation, or Qvxxqz Blorpblorp to watch Vyuha refuse to guess)
